Research Of Objects Association In Non-overlapping Multi-camera Surveillance System

With the development of intelligent video surveillance system,the research of objects association in non-overlapping multi-camera has become a key research direction in the field.Objects association is a recognition task in which one matc hes the individuals across cameras in disjoint views,which is also known as person re-identification.The complex variations in illumination,viewpoint and posture across different camera views can cause large appearance variance,which makes objects association still a challenging problem.To address these problems,we conduct a further study on video-based objects association and image-based objects association in this thesis,respectively.The main work and innovations are as follows :1.We design a re-identification method based on appearance model and spatio-temporal model for video-based objects association.We firstly use color normalization to reduce the influence of illumination difference between the cameras and extract the appearance features of the object.Then we utilize the simple accumulation of the object appearance similarity under different transfer times to learn the topological relationship of the camera network in an unsupervised way,these spatio-temporal features are irrelevant to the camera's illumination,environment and viewpoint.Finally,we combine them with the appearance features of the object to gain more representive features to achieve a higher matching rate.2.We design a re-identification method based on deep feature representation and multiple metric ensembles for image-based objects association.Firstly,we jointly use the various benchmark datasets to train a common Convolutional Neural Network(CNN)which is employed to extract more deep and intrinsic features of the object,and then we exploit multiple metric ensembles to measure the similarity between the features in a more comprehensive way,which can improve the discrimination of the distance mesure.Experimental results demonstrate that our method can effectively improve the matching rate of objects association,and achieve excellent performance on multiple public benchmark datasets.